WeltGet startedLearn in a notebook
Learn in a notebook
Work through Welt's workflows as notebooks on real research data. Start with classification.
Each notebook downloads real research data, prints the table and explains what happens next. You don't need a helper file or a testing context. The first cell installs the reader; the next cell connects your workspace.
For each notebook, choose Open in Colab to run it, download the .ipynb file for
Jupyter, or choose View to read its formatted cells here. All three use the same
notebook.
Notebooks use workspace credits
Modelling and research downloads may consume credits. Check Workspace → Billing for available credits and spend caps before you run the notebooks. See usage and recovery.
Classification#
Download Iris, inspect it, split train and test rows, fit, predict, then read accuracy and a per-class report.
Regression#
Download Yacht, inspect it, split train and test rows, fit, predict, then read MAE, RMSE and R².
Async jobs#
The Iris workflow with durable submission and async waiting.
Artifacts and usage#
Evaluate Iris, reopen the same predictor and inspect its operation events.
Compare models#
Hold out Iris test rows, cross-validate models on the rest, read the leaderboard and the report, then fit the chosen model and check it once on the test rows.
Research datasets#
Search the catalogue, inspect a fixed eligible asset and complete your first classification.
Causal discovery#
Inspect real Iris measurements and the graph structure Welt's native model proposes.
Before you run#
- Use a fresh Python 3.11–3.13 environment with
welt-client[parquet]. - Outputs in the notebook sources are cleared. Your own run shows actual data and measurements.
- Keep credentials in
WELT_API_KEY, never in saved notebook cells. - Every modelling call goes to Welt. The private training data, predictors and results you create persist.
About the causal example
It has no ground-truth graph and makes no causal-accuracy claim. Native scores are
not calibrated confidence. To continue in ergodic, install welt-client[ergodic] and
call result.to_ergodic() (declared DAGs only); see
Convert to an Ergodic graph.
How are these notebooks checked?
Local release checks patch transport outside the public notebook and execute using authorized SHA-pinned research files. Public CI validates sources and contracts without downloading research bytes or using account credentials. Fixture scores never measure model quality; actual hosted acceptance is recorded separately.
Run in Colab#
Choose Open in Colab and run the cells in order. Each notebook starts with:
!pip install "welt-client[parquet]"
If Colab asks you to restart the session, restart it after installation. Then continue with the connection cell: open its printed verification link and approve your workspace. View pages let you read the same cells without running them.
Next steps#
- Your first prediction the same classification and regression steps as a page.
- Use your own table run the workflow on your data.
- Connect your workspace connection, saved access and remote use.